Overview
- Moonshot published the full Kimi K3 weights and a technical report on Hugging Face and GitHub under a custom Kimi K3 licence, completing the company’s public rollout on July 27.
- Kimi K3 is a 2.8 trillion‑parameter Mixture‑of‑Experts model with a one‑million‑token context window that activates about 104 billion parameters during inference and ships in MXFP formats with optimized attention kernels.
- The raw weights are roughly 1.4 terabytes and practical self‑hosting requires many high‑end GPUs, large networking and storage, so third‑party hosts such as Telnyx have begun offering OpenAI‑compatible inference and published pricing.
- Moonshot’s benchmarks and early community tests place K3 near frontier closed models on coding and agentic tasks while the company claims much lower per‑token running costs, especially when prompt caching helps.
- The release has sparked allegations from Anthropic and U.S. officials that Moonshot used distilled outputs from Western models, a denial from Moonshot, and renewed debate over export controls and limits on open‑weight distributions.